Skip to content

Local contributor workbench

The local workbench is a small, offline-first contribution surface alongside the deterministic SPC walkthrough. It accepts CSV and JSONL, shows inferred mappings, validates records, and lets a contributor inspect provenance, structures, graph readiness, slices, audited baseline rankings, and normalized export.

Limits and lifecycle

  • CSV and JSONL only; 5 MiB and 5,000 rows maximum.
  • At most 100 individual issues are returned; aggregate counts remain available.
  • Strict rejection is the default. Skipping invalid rows must be explicitly selected and marks the dataset degraded.
  • Canonical normalized JSONL bytes are kept in memory. The registry holds at most eight datasets and 32 MiB, evicting least-recently-used entries as needed.
  • Only the selected dataset is deserialized into a MaterialStore.
  • Imported content is not sent externally or written to disk. Reset deletes the active imported entry and returns to the bundled demo.

Dataset IDs are derived from the SHA-256 of canonical normalized JSONL. Export responses include dataset ID, checksum, and record-count headers.

CSV mapping

Map an identity column, formula, optional structure JSON, optional upstream source ID, and any number of property columns. Each property mapping has a canonical name, optional unit, source, and method. Unknown units are retained verbatim with a warning; mixed methods remain visible per property.

JSONL records may already use the public Material shape. Each accepted row receives local-file provenance containing only the safe basename, input checksum, and row number.

API

The UI uses /datasets/inspect, /datasets/import, /datasets, dataset detail/delete/export, and /datasets/{id}/slices/preview. Existing materials, search, graph-summary, audited-ranking, and ASE smoke-test routes accept an optional dataset_id; omitting it always selects the bundled demo.

The cached CHGNet AlN reference is never presented as applicable to imported data.